Tron-1B
Tron-1B answers typed questions about text or JSON in a single pass: choose one option, rate on a scale, or answer yes/no, with calibrated probabilities. It is built for the fast decisions around an AI application: routing, triage, safety screening, intent detection, and workflow automation.
- Accurate: beats Jev 1.13.0 on all four of its published benchmarks (below).
- Fast: 16.7 ms per decision (p50) on one GPU.
- Calibrated: its confidence scores track how often it is right, so you can gate actions on them.
- Any question, at request time: you write the question and the options; no retraining for new label sets.
Try it
The quickstart notebook runs on Colab's free T4 GPU: routing, safety screening, all 77 Banking77 intents, a speed test, and a form for your own questions.
Quickstart
pip install https://huggingface.co/SamCodeManMk2/tron-1b/resolve/main/troncore-1.0.0-py3-none-any.whl
from troncore import Engine
eng = Engine("SamCodeManMk2/tron-1b") # downloads once, then runs locally (GPU if available)
answers = eng.decide(
{"subject": "Duplicate charge on invoice #4411",
"body": "We were billed twice for March. Refund the duplicate today or we cancel our plan."},
{
"department": {"type": "choice", "instructions": "Which team should handle this?",
"criteria": {"billing": "invoices, payments, refunds", "technical": "bugs, outages",
"sales": "pricing, new contracts", "other": "anything else"}},
"urgency": {"type": "score", "instructions": "How urgent is this?",
"criteria": ["not urgent", "this week", "today", "critical"]},
"churn_risk": {"type": "yesno", "instructions": "Does the customer threaten to cancel?"},
},
)
answers["department"]["choice"] # 'billing'
answers["churn_risk"]["probability"] # P(yes)
answers["department"]["confidence"] # calibrated confidence of the top option
Every answer includes probabilities, confidence, margin (top minus second) and entropy. Pass
min_confidence=0.8 to get abstain: true on uncertain answers, so they can be routed to a person or a larger model.
Questions with more than 128 options are handled automatically by scoring them in rounds. Long inputs can be read
in sliding windows with windows="all".
Question types
| type | options | answer |
|---|---|---|
choice | criteria: list of labels, or dict label → description | choice + a probability per label |
score | criteria: ordered list of levels | score (expected level), level, a probability per level |
yesno (alias noul) | fixed no / yes | answer (bool) + probability of yes |
Results
| Benchmark | Tron-1B | Jev 1.13.0 |
|---|---|---|
| Banking77 (intent, 77 labels)¹ | 94.0 | 87.0 |
| AG News (topic) | 93.9 | 91.0 |
| typed-decisions (2,000 business decisions) | 79.6 | 72.7 |
| DAIR emotion | 92.9 | 48.0 |
| Latency, p50 per decision² | 16.7 ms | 236–276 ms |
Tasks Tron-1B never trained on (whole task families held out of training):
| Task | Accuracy |
|---|---|
| IMDB sentiment | 94.7 |
| Prompt-injection detection (deepset) | 90.5 |
| MASSIVE intent (59 labels) | 88.4 |
| XNLI entailment (English) | 86.7 |
| BoolQ reading comprehension | 79.2 |
| CLINC intent (151 labels) | 62.8 |
| SST-5 graded sentiment | 54.8 |
¹ Jev's published Banking77 result used 72 labels; Tron's uses all 77. Jev figures are its published numbers, not re-run by us. ² Tron: one question on one NVIDIA GB10 GPU. Jev: independently measured p50 through its hosted API, which includes network time.
Tron-1B was trained on the training splits of Banking77, AG News and DAIR emotion and evaluated on their official test splits. All results measured 2026-09-28 with troncore 1.0.
How it works
Tron-1B is a 1.1B-parameter bidirectional encoder (initialised from ettin-encoder-1b) with a decision head. For each question, the question, the input and every option are encoded together. Each option is pooled into a vector, and a small attention layer compares the options with each other and with the question before they are scored. This is what lets it separate close labels such as "card not arrived" and "card delivery estimate". Probabilities are calibrated per question type and option count.
It was trained on about 1.2 million typed decisions built from public classification, entailment, safety, routing, reasoning, preference and business-workflow datasets.
Limitations
- It decides; it does not write. Answers are always one of the options you give it.
- Very large unseen label sets are its weakest area (62.8% on CLINC's 151 intents zero-shot). For big label sets, short descriptive label names help, and so does a few hundred examples of fine-tuning.
- Long inputs: accuracy is best under about 2,000 tokens of input. Send the relevant section rather than a whole
document, or use
windows="all". - English first. It handles other languages, but was mostly trained and evaluated in English.
- Not a safety guarantee. Use its safety and injection judgements as one layer of defence, with confidence gating, not as the only one.
License
The model weights are released under CC BY-NC 4.0: free to use,
share and adapt for non-commercial purposes, with attribution. The troncore runtime is Apache-2.0. The base
encoder is MIT-licensed; see NOTICE.